The blog
Practical AI, for businesses that have real work to do.
Most writing about AI is either a product launch or a prediction. This is neither. It is what Claude is actually being used for inside ordinary businesses — the invoice pile, the support inbox, the contract review that eats a week — what it costs, and how to find someone who has done it before.
Start here
How to hire a Claude consultant
How to tell a real practitioner from someone who watched a course, what an engagement costs, and the questions worth asking before you commit any money.
ReadAutomating a process
The expensive, boring work happening by hand right now — invoices, inboxes, documents, data entry — and what it takes to automate it properly.
How to query your company data in plain English
Translating a question into a query is largely solved. Deciding what your business means by 'active customer' is not, and that is the project.
ReadAI contract review: what it can and can't do
Extraction and playbook comparison work well. Judgement about acceptable risk does not transfer. Why the playbook is the project, not the technology.
ReadHow to automate customer support emails
Drafting and triage before auto-sending. What to automate first, where a human stays, and why deflection is the wrong number to optimise.
ReadHow to automate invoice processing with AI
Anthropic ships invoice tracking and month-end close workflows out of the box. Here is what they do, the four things they cannot know about your business, and what it takes to close that.
ReadExtracting data from PDFs and scanned documents with AI
Template OCR works for your top twenty suppliers and fails on the hundred below them, which is where the cost is. What changes when the system reads instead of matching.
ReadBy industry
What Claude is actually being used for, sector by sector.
AI for logistics and freight companies
Documentation, exception handling and status chasing are where freight margin disappears. What responds well, and why the awkward system is always the TMS.
ReadAI for law firms
Review, disclosure and first drafts are where the hours go. Why citation, privilege and the billing model decide whether a deployment survives.
ReadAI for accounting firms
Templates now exist for ledger reconciliation, month-end close and statement audit. The gap is your clients' chart of accounts, your review policy, and who signs.
ReadThe Claude platform
The independent implementation record: what Anthropic actually shipped, what it changes for people building on it, and when we last checked. Sourced from release notes and docs, not press coverage.
Claude Academy: what it is, and how it differs from Claude certification
Anthropic’s learning hub has a new home and a new name. What is in it, whether it costs anything — and the distinction that matters on a CV: a completion badge is training evidence, a certification is a proctored exam.
ReadClaude’s text watermark: how it works and what it proves
An invisible statistical watermark now rides in Claude’s word choices, with C2PA metadata on generated image files. What it marks, what it cannot prove, and why there is no opt-out.
ReadThe Claude enterprise control plane: compliance API, inference hooks and agent budgets
Session transcripts from users' own machines, prompt-level allow/deny against your security server, and hard spend caps on agent sessions. What each control does and where it stops.
ReadHiring an AI specialist
What it costs, how to scope it, how to tell a real practitioner from someone who watched a course, and what belongs in the contract.
How to vet an AI consultant
Four checks that take an hour: verify the credential at source, ask what broke, read the proposal for assumptions, and give the same brief to two people.
ReadHow to write a brief for an AI project
A brief that names a technology has already chosen the answer. What to include instead, and the one constraint that saves the most time when stated up front.
ReadHow much does AI implementation cost?
The four things that actually drive the price, why a fixed quote given without questions is a warning sign, and how to size the budget from your own numbers.
ReadWhen it goes wrong
Pilots that never reached production, chatbots that invent things, tools nobody uses. What went wrong and what fixes it.
Getting an AI proof-of-concept into production
The demo worked and the rollout stalled. The five things a proof of concept is allowed to skip, and why each one is most of the remaining work.
ReadWhy nobody uses the AI tool you bought
It works, it is paid for, and usage collapsed after week three. The four reasons adoption fails, and why one bad answer costs more than ten good ones earn.
ReadHow to stop an AI chatbot hallucinating
Telling a model not to make things up does not work. Constraining its sources, forcing citation and designing a good refusal does — and all three are measurable.
ReadCompanies rehiring after AI layoffs: what went wrong
The sequence is consistent and avoidable: cost reduction chosen as the goal, headcount cut on a demo, and the checking work discovered afterwards.
ReadWhy do AI projects fail?
Four failure modes, none of them the model: a problem nobody scoped, data nobody could reach, no definition of working, and no owner afterwards.
ReadBuilding a practice
For certified practitioners: pricing, proposals, first clients, and making a twelve-month credential pay for itself.
How to get your first AI consulting client
The first client almost never comes from marketing. It comes from a specific problem, in a network you already have, described better than the person who has it.
ReadHow much do AI consultants charge?
For practitioners setting a rate: why the day rate caps you, how to price discovery as its own product, and what to charge for the second engagement.
ReadIs the Claude certification worth it?
Four credentials, twelve-month validity, and a very different answer depending on whether you are employed or selling your own time.
ReadClaude Certified Architect vs Developer vs Associate
Four credentials at three price points. Match it to the work you intend to sell rather than to seniority, and check what a buyer will actually recognise.
ReadRisk & governance
Data access, contracts, model training, liability — the questions that stop a project before cost ever comes up.
Is it safe to give an AI consultant access to company data?
The controls are ordinary supplier-access controls plus three questions specific to AI: training, retention and residency. Start with a redacted sample, not production.
ReadWho is liable when AI gets it wrong?
Liability for an AI system's output usually sits with the business that deployed it. What the contract should say, and why the audit trail matters more than the accuracy figure.
ReadWhat to include in an AI project contract
Seven clauses that decide how the engagement ends. The acceptance criteria matter most, and they are the ones most often left as an intention.
ReadWhere AI fits
An eight-part series for people who run businesses, or work in them, and have been asked “where can we use AI?” without being given any way to answer. One test, applied to the eight places paperwork lives, followed through one real company — including the parts that went wrong. See docs/WHERE-AI-FITS.md.
Nothing matches that yet.